WO2017158058A1 - Procédé de classification de cas uniques/rares par renforcement de l'apprentissage dans des réseaux neuronaux - Google Patents

Procédé de classification de cas uniques/rares par renforcement de l'apprentissage dans des réseaux neuronaux Download PDF

Info

Publication number
WO2017158058A1
WO2017158058A1 PCT/EP2017/056172 EP2017056172W WO2017158058A1 WO 2017158058 A1 WO2017158058 A1 WO 2017158058A1 EP 2017056172 W EP2017056172 W EP 2017056172W WO 2017158058 A1 WO2017158058 A1 WO 2017158058A1
Authority
WO
WIPO (PCT)
Prior art keywords
data
label
dnn
deep neural
labels
Prior art date
Application number
PCT/EP2017/056172
Other languages
English (en)
Inventor
Dzmitry Tsishkou
Rémy Bendahan
Original Assignee
Imra Europe Sas
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Imra Europe Sas filed Critical Imra Europe Sas
Priority to EP17710934.5A priority Critical patent/EP3430576A1/fr
Priority to JP2018549257A priority patent/JP6983800B2/ja
Priority to US16/085,989 priority patent/US10776664B2/en
Publication of WO2017158058A1 publication Critical patent/WO2017158058A1/fr

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05D—SYSTEMS FOR CONTROLLING OR REGULATING NON-ELECTRIC VARIABLES
    • G05D1/00—Control of position, course or altitude of land, water, air, or space vehicles, e.g. automatic pilot
    • G05D1/0088—Control of position, course or altitude of land, water, air, or space vehicles, e.g. automatic pilot characterized by the autonomous decision making process, e.g. artificial intelligence, predefined behaviours
    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05D—SYSTEMS FOR CONTROLLING OR REGULATING NON-ELECTRIC VARIABLES
    • G05D1/00—Control of position, course or altitude of land, water, air, or space vehicles, e.g. automatic pilot
    • G05D1/02—Control of position or course in two dimensions
    • G05D1/021—Control of position or course in two dimensions specially adapted to land vehicles
    • G05D1/0231—Control of position or course in two dimensions specially adapted to land vehicles using optical position detecting means
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/211—Selection of the most significant subset of features
    • G06F18/2113—Selection of the most significant subset of features by ranking or filtering the set of features, e.g. using a measure of variance or of feature cross-correlation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G06F18/2148—Generating training patterns; Bootstrap methods, e.g. bagging or boosting characterised by the process organisation or structure, e.g. boosting cascade
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G06F18/2155—Generating training patterns; Bootstrap methods, e.g. bagging or boosting characterised by the incorporation of unlabelled data, e.g. multiple instance learning [MIL], semi-supervised techniques using expectation-maximisation [EM] or naïve labelling
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G06N20/20—Ensemble learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/0895—Weakly supervised learning, e.g. semi-supervised or self-supervised learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/096—Transfer learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/50—Context or environment of the image
    • G06V20/56—Context or environment of the image exterior to a vehicle by using sensors mounted on the vehicle

Definitions

  • the present invention relates to machine learning techniques and more particularly to deep neural networks (DNN) such as deep convolutional neural networks (CNN).
  • DNN deep neural networks
  • CNN deep convolutional neural networks
  • the present invention relates to a method of classification of unique/rare cases by reinforcement learning in neural networks. Such rare cases could be defined as a situation or scenario that could be significantly different from all previously learned data and have minor chances to occur in everyday situations.
  • Such method is useful especially in the field of human-assisted or autonomous vehicles using a camera or a depth sensor such as a Lidar sensor, for obstacle detection and avoidance to navigate safely through environments.
  • Generative Adversarial Networks discloses the use of an adversarial process of simultaneous training of both generative and discriminative models.
  • Generative model should model data distribution from previously labelled data samples in order to be able to synthesize new data sample.
  • the discriminative model must be trained to estimate the probability of this new sample being from the training data set rather than generated by the model. This repetitive process corresponds to a minimax two-player game, where reinforcement learning leads to an improvement of the discriminative model.
  • the present invention aims to address the above mentioned drawbacks of the prior art, and to propose according to a first aspect to overcome this problem by using a deep neural network combining feature spaces as well as decision spaces in which both visually similar and visually non-similar misclassification data samples could be hosted and classified.
  • a first aspect of the invention relates to an image processing method which comprises the steps of: - capturing image data as original data (Data-A);
  • - reinforcing deep neural network learning capacity to classify rare cases amongst captured image data, comprising the sub-steps of: - training a first deep neural network (DNN-A) used to classify generic cases of said original data (Data-A) into specified labels (Label-A);
  • Label-B spatial-probabilistic labels
  • DNN-C deep neural network
  • DNN-D deep neural network
  • the image processing method according to the first aspect comprises more particularly the steps of:
  • DNN-A deep neural networks
  • Label-A a first deep neural networks
  • HEAT-B - localizing discriminative class-specific features
  • Label-B mapping the discriminative class-specific features on the original data as spatial- probabilistic labels
  • DNN-C deep neural network
  • DNN-D combined deep neural network
  • Data-C unlabelled data
  • Label-A+B * classifying the unlabelled data into the primary combined specified and spatial-probabilistic labels
  • Label-C mapping the primary combined specified and spatial-probabilistic labels
  • Label-C secondary combined specified and spatial-probabilistic misclassification labels
  • Such image processing method uses a deep neural network combining universal and class-specific spaces in which both visually similar and visually non-similar misclassification data samples can be easily hosted and classified. More particularly, in the first deep neural networks, trained feature space filters non-discriminative information, which serves as a pre-filter to find the most discriminative information on the next step.
  • the localizing and mapping step adds a spatial and probabilistic dimension of labelling information associated with original images.
  • this information is used to make a universal feature space based on the discriminative class-specific features, by removing class-specific information. Both the class-specific feature space and the universal feature space are complementary to handle efficiently visually similar and visually non-similar misclassification data samples.
  • the method forms new misclassification labels around each group of misclassification errors hosted in the new primary combined class-specific and universal feature/decision spaces.
  • large source of unlabelled data which could be any kind of natural images/videos
  • a penalty matrix which is used to control distribution of unlabelled data samples used for training, in order to re-balance future misclassification mistakes for each rare/unique case according to prior estimates of application-level risk (defined as penalty for misclassification between each pair of labels in the penalty matrix).
  • penalty matrix could be defined automatically prior to the training stage or re-defined by human experts at each iteration of the reinforcement loop.
  • the first, second and combined deep neural networks are convolutional neural networks.
  • the different deep neural networks allow to use bi-dimensional inputs.
  • the localizing step of discriminative class-specific features consists in:
  • Label-A - ranking randomly extracted patches by their statistical popularity among all patches of a same specified label
  • Such steps of localizing and mapping discriminative class-specific features reformulate way of defining regularization process for feature/decision space creation by deep neural networks, which are using a loss function working with specified labels into the process that uses spatial- probabilistic labels related to spatial locations and corresponding probability of the most discriminant features (i.e. non-class specific) in the form of a heat map.
  • feature/decision space creation process allows to train feature/decision space of universal features, which are complementary to class-specific features. Combination of both feature spaces makes then possible to train a deep neural network that can confidently discriminate between various types of misclassification errors since it can also discriminate by universal features, which were previously indistinctly merged in class-specific only feature/decision spaces.
  • the randomly extracted patches are with high level of activation of the original data processed through said at least one class- specific feature spaces; bag of visual words features are done for highly ranked patches and the extracted variable importance measures are using the random forest tree classification to localize highly discriminative class- specific features.
  • the training step of the second deep neural network (DNN-C) further includes:
  • Label-B - resizing spatial-probabilistic labels
  • Label-A - estimating localization and probability of the discriminative class- specific features independently from specified labels
  • Such training step of the second deep neural network ensures to resize the heat map received from the previous step, to learn the universal feature/decision spaces corresponding to the discriminative class-specific features and classify rare cases into the resized spatial-probabilistic labels.
  • the training step of the first combined deep neural network (DNN-D) further includes:
  • Label-A+B * a combined loss function of the universal loss function of the second deep neural network and a class-specific loss function of the first deep neural network to provide feedback for original data belonging to any label out of the primary combined specified and spatial-probabilistic labels
  • Combination or fusion of universal and class-specific feature/decision spaces allows training such combined deep neural network with any kind labelled or unlabelled data.
  • the learning rate of the universal feature space is smaller than the learning rate of the class-specific feature space.
  • the training step of the second combined deep neural network (DNN-F) further includes:
  • the method defines new misclassification labels for the loss function of the second combined deep neural network which then will better learn feature/decision space in locations of unseen/rare misclassification mistakes, as sub-labelling of mistakes related to original specified labels can be done in the combined universal/class-specific feature/decision spaces from the first combined deep neural network. Further, it can then use unlabelled data to feed the first combined deep neural network (DNN-D) and collect enough data samples for each newly defined sub-labels of mistakes. Then by training another (second) deep neural network (DNN-F) that was preferably pre-trained from the first combined deep neural network (DNN-D) on newly collected data samples from originally unlabelled data source, combined feature/decision space in previously unseen/rare locations can be efficiently improved.
  • DNN-F second deep neural network
  • the method further comprises the step of:
  • DNN-G third combined deep neural network
  • the training step of the third combined deep neural network (DNN-G) further includes:
  • Loss-A modifying a loss function (Loss-A) of the third combined deep neural network or batches of unlabelled data by using more data samples for the rare cases, with higher penalty on misclassification based on the penalty matrix.
  • the training step of the third combined deep neural network uses knowledge transferred from the trained second combined deep neural network (DNN-F) which could classify input unlabelled data into sub-labels of mistakes.
  • This third combined deep neural network has the same input and out configuration as the initial deep neural network (DNN-A) and thus allows making a big reinforcement loop.
  • the reinforcement method should keep further improvements of the final network (DNN-G), since during each iteration new data samples from unlabelled data would be learned.
  • mapping step of primary combined labels (Label- A+B * ) into secondary combined misclassification labels (Label-C) together with the training steps of both the second and third combined deep neural networks form a small reinforcement loop; and the steps from the training step of the first deep neural network to the training step of the third combined deep neural network form a big reinforcement loop.
  • DNN-A initial deep neural network
  • DNN-G more accurate and reliable global network
  • the present invention In contrast with the "Distilling the Knowledge in a Neural Network" publication, in the present invention, it is proposed to use unlabelled data to construct feature/decision spaces not only near previously learned data samples, but everywhere else by focusing more on unknown locations/areas with high risk of misclassification defined by the penalty matrix. In such way, the present invention provides much large and denser coverage of feature/decision space for previously rare/unique/unseen cases, i.e. data samples, which leads to more accurate and robust classification.
  • the problem of data augmentation is addressed by using unlabelled data, which has potentially orders of magnitude with higher variability compared to data augmentation from existing labelled data samples), and augmenting feature/decision spaces that allows to specifically search for previously unseen/rare cases that might have high risk of misclassification. Therefore with the present invention, it is possible to collect much more adversarial data samples of much higher variability, which leads to construction of better feature/decision spaces and finally to more accurate classification and lower level of false alarms.
  • the invention relates to a vehicle comprising:
  • a path capturing unit arranged to capture and convert portions of a followed path seen at least from a driver's point of view into a series of digital files, when the vehicle is driven
  • a processing unit hosting a deep neural network, arranged to classify rare cases based on the series of digital files according to the above image processing method
  • a display unit arranged to display an information related to the classified rare cases
  • an autonomous driving unit arranged to control the vehicle
  • a decision unit arranged to activate at least one of the display unit and the autonomous driving unit depending on the classified rare cases.
  • the processing unit is arranged to be updated online via cloud-computing processing computers or offline in a service mode integrating the latest results of the distributed reinforcement loop process.
  • FIG. 1 represents a method to reinforce deep neural network learning capacity to classify rare cases according to a preferred embodiment of the present invention
  • FIG. 2-8 represents stages A-G according to a preferred embodiment of the invention
  • FIG. 9 represents a vehicle equipped with the necessary units to implement the method according to the invention.
  • General summary Figure 1 represents a method to reinforce deep neural network (DNN), such as convolutional neural network, learning capacity to classify rare cases according to a preferred embodiment of the present invention.
  • DNN deep neural network
  • the whole DNN training process comprises several stages A-G forming a big reinforcement loop (Stage I). Within the big reinforcement loop, Stages E-G forms a small reinforcement loop (Stage H).
  • the main goal of the whole DNN is to reinforce training of DNNs using the following algorithm.
  • the algorithm trains a state of the art DNN-A using Data-A and Label-A.
  • Stages B-D big reinforcement loop
  • the algorithm connects to hidden layers of DNNs, reinterprets signal activations as universal features, then makes combined feature/decision spaces (primary spaces) that can properly host/classify rare cases followed by a small reinforcement loop.
  • Stages E-G small reinforcing loop
  • the algorithm maps the primary spaces into secondary feature/decision spaces and fills them up with unlabelled data; such secondary spaces have enough data samples and layers to learn rare cases as good as generic cases.
  • a penalty matrix gives priors on how feature space should be constructed. Priors refer generically to the beliefs an agent (i.e. the programmer) holds regarding a fact, hypothesis or consequence.
  • the secondary feature/decision spaces are warped down to match the state of the art DNN-A's inputs and outputs. During this stages one can't assess quality of the mapping directly, however at least two indirect methods could be used. Firstly, accurate evaluation of DNN-C can be done by further splitting its training data, where ground truth is available, into training/test batches.
  • unlabelled data that may be used do not need to cope with quality requirements. Indeed, any kind of images could be used as unlabelled data even synthetic or randomly generated.
  • everyone could get billions of images to obtain unlabelled data.
  • Closer the domain of unlabelled images is to the target domain, higher overlap between newly learned features/decision spaces of labelled/unlabelled data would be.
  • any natural/synthetic outdoor or indoor images could be used.
  • image quality or resolution is not an important factor in case of usage of millions of image samples since most of the visual features are repeatable amongst images.
  • the achieved goal of the whole DNN is to reinforce DNN learning capacity to improve classification accuracy for rare/unique cases according to priors on risk-of-misclassification defined as a penalty matrix (based on general knowledge).
  • DNN One example of DNN that can be used is lnception-V3, more details being given at: http://arxiv.orq/pdf/1512.00567v3.pdf. Stage A
  • Figure 2 represents Stage A of training of a first deep neural network (DNN-A). Supplied data are images (Data-A) and tags (Label-A), where images of x * y are tagged as instances of 1 :N classes.
  • Data-A images
  • Label-A tags
  • the algorithm trains DNN-A, where both feature space (Conv-A) related to convolutional layers and decision space (FC-A) related to fully connected layers shall be constructed via back-propagation of feedback generated by a loss function (Loss-A) given forward signal from mini-batch of images with tags.
  • Conv-A feature space
  • FC-A decision space
  • Loss-A loss function
  • Stage A output includes decision (FC-A) and feature (Conv-A) spaces, which are used by DNN-A to automatically classify Data-A like image instances into Label-A specified tags.
  • Stage A goal is that trained feature/decision spaces filter non- discriminative information within Data-A, which serve as pre-filter to find the most discriminative information on the next step.
  • Stage B Figure 3 represents Stage B of "Plug-in" images and tags and
  • Supplied data are the same as stage A, namely images (Data-A) and tags (Label-A), where images of x * y are tagged as instances of 1 :N classes, and the images (data-A) passed through convolutional layers, while keeping only strong activations as input for stage B.
  • the algorithm finds locations of class-specific highly discriminative features by : - random extraction of patches (Ext-B) with high level of activations of image signal processed through convolutional layer of DNN-A;
  • RFT-B random forest tree classification of bag of (visual) words features (BOW-B) converted data
  • IMP-B - variable importance
  • RFT-B random forest tree
  • HEAT-B HEAT-B
  • IMP-B most discriminative patches
  • Label- B mapping on original data as new spatial-probabilistic labels
  • Stage B output includes heat maps of the most discriminative class-specific features linked with original image data.
  • Stage B goal is that the heat maps add new dimension of labelling information associated with original images (spatial-probabilistic + specified class vs. specified class). This information will be used in the next stage to make universal feature space, by removing class-specific information.
  • the random extraction patches can be done from the activation map of a convolutional layer of the deep neural network (e.g. like layer N of Conv-A).
  • Explanation about activation maps can be obtained for example from http://cs231 n.qithub.io/understandinqcnn/.
  • Each convolutional layer would be used separately for extraction of random patches, so we would make patch ranking for layer 1 , layer 2, ... and later we could select the best result to make a fusion.
  • the activation maps of convolutional layers are sparse, e.g. 96 filters - 96 activation maps, where each filter is tuned to respond only to specific signal
  • high level of activation corresponds to the filter's response level to the input signal (i.e. original image for the first layer, or output of the previous hidden layer) above some threshold.
  • an example in the present case is the following.
  • BOW concept as feature vector
  • Each visual word is considered as a variable while learning random forest tree classifier, variable importance is computed to estimate how classification performance of the forest tree classifier, that uses 5% of total visual words (since we keep only high discriminative patches ⁇ 80% of total population of patches covered) depends on contribution of this visual word.
  • variable importance is computed to estimate how classification performance of the forest tree classifier, that uses 5% of total visual words (since we keep only high discriminative patches ⁇ 80% of total population of patches covered) depends on contribution of this visual word.
  • any of 5% of visual words can be more or less popular according to the patch ranking, but their popularity is measured only within its own class samples, so by taking 5% of highly ranked patches we ensure to cover 80% of total samples. Then by computing variable importance, we understand on which of these samples random forest tree classifier relies to discriminate between two classes. Indeed, if patch is popular for many classes it would have low variable importance, only patches which are popular within their own class and not popular within other classes would have high importance, since random forest classifier would rely on them.
  • heatmap shows most discriminative patches locations on a map in some pre-defined coloured areas (e.g. red), which are spatial- probabilistic labels.
  • each heatmap is normalized between min/max values to fit 0:256 range.
  • Such heatmap is a Label-B sample for corresponding image.
  • Stage C Figure 4 represents Stage C of topological mapping of a second deep neural network (DNN-C).
  • Supplied data are images (Data-A) and new spatial-probabilistic tags (Label-B, heatmaps x * y).
  • the algorithm trains DNN-C by: - resizing x * y heatmaps from Label-B to M classes, where 1/sqrt(M) is the resolution of spatial-probabilistic labels (Label-B);
  • Stage C output includes decision (FC-C) and feature spaces (Conv-C), which are used by DNN-C to automatically map Data-A like image instances into resized spatial-probabilistic labels (Label-B * ; heatmaps 1 :M) based on universal feature space.
  • Stage C goal is that the universal feature space is complimentary to class-specific feature space (obtained on stage A), so that later fusion of two spaces should improve DNN's learning capacity.
  • the resolution of spatial-probabilistic labels is defined by dimension of activation map of convolutional layer, where patches are computed. For example we can use activation maps of a first convolutional layer of 56x56), thus M is smaller or equal to 56x56.
  • the spatial part is linked to the resolution of the heatmap and defines number of classes M and the probabilistic part defines distribution of samples of each of 1 :M classes during training stage, such that samples with labels from Label-B which have higher probability would be overrepresented during training compared to the lower probable samples.
  • training stage of DNN is divided into series of smaller stages, where at each smaller stage only a tiny randomly picked portion of data (i.e. mini-batch) is used to compute loss and do back-propagation.
  • mini-batch can be composed of 200 randomly chosen samples out of 1 .000.000 or more.
  • mini-batch can be composed of 200 randomly chosen samples out of 1 .000.000 or more.
  • universal feature they are features not linked to a specific class and can be shared among different objects of different classes.
  • Stage C we force DNN-C to learn to localize discriminative part of objects and estimate their probability (e.g. how discriminative it is).
  • Figure 5 represents Stage D of primary combined feature/decision space in a primary combined deep neural network (DNN-D).
  • Supplied data are images (Data-A) and tags (Label-A; 1 :N classes) and resized spatial-probabilistic tags (Label-B * ; 1 :M classes).
  • the algorithm trains DNN-D by:
  • FC-C ⁇ A ⁇ D - repeating previous step for decision space to learn combined decision
  • FC-C ⁇ A ⁇ D - using combined loss function of Loss-C and Loss-A to provide feedback for each image signal belonging to one of classes out of M * N.
  • Stage D output includes decision (FC-C- A- D) and feature (Conv- C- A- D) spaces, which are used by DNN-D to automatically map Data-A like image instances into combined universal/class-specific feature space (Label-A+B called "Primary").
  • Stage D goal is to combine both universal and class-specific feature spaces, so that all rare cases can automatically be clustered.
  • Many rare cases which have same class-specific classes (1 :N) according to DNN-A, will be repeatedly subdivided into universal/class-specific sub-classes (1 :M * N) by DNN-D. Further, even if rare cases present the same class-specific features, they could be distinguished by their universal features which would be different from one rare case to another and then help to better trained on these rare cases and therefore increase accuracy and reduce false alarms.
  • DNN-D we use transfer learning from DNN-A and DNN-C.
  • DNN-D we use transfer learning from DNN-A and DNN-C.
  • DNN-D we use transfer learning from DNN-A and DNN-C.
  • DNN-D we use transfer learning from DNN-A and DNN-C.
  • DNN-D we use transfer learning from DNN-A and DNN-C.
  • DNN-D we use transfer learning from DNN-A and DNN-C.
  • M * N instead of either 1 :N for DNN-A or 1 :M for DNN-C.
  • DNN-A, DNN-C and DNN-D we could define DNN-A, DNN-C and DNN-D to have same configuration for all layers expect final fully connected layer.
  • the learning rates it defines how important would be changes at each back-propagation stage to the DNN for each layer. For instance, in case of high learning rate, backpropagation would significantly modify DNN connections on each training iteration. In our present case, we can consider for example that first hidden layers are mostly related to DNN-C, while deeper hidden layers are mostly related to DNN-A. If we fix small learning rate for first hidden layers, they would be modified significantly less that other part of the network, so we could consider that most of the universal features would be kept. But since the learning rate is not zero, we still allow these universal features to be slightly adapted to prevent back-propagation from saturation.
  • DNN-C is trained to classify input image into 1 :M classes
  • DNN-A is trained to classify input image into 1 :N classes.
  • DNN-D with DNN-C, it would already have features required to predict labels 1 :M, knowing that this knowledge was transferred.
  • Figure 6 represents Stage E of mapping unlabelled data to secondary classes. Supplied data are unlabelled images (Data-C) on a much higher scale than images (Data-A).
  • Stage E output includes assigned tags in Secondary space of 1 :M * N classes related to Primary space (Label-C; 1 :M * N classes). These assigned tags (Label-C) are also called secondary combined specified and spatial- probabilistic misclassification labels.
  • Stage E goal is to always find enough data samples related each "Secondary" version of rare case by using sufficiently large source of unlabelled data (Data-C), so that DNN-D could be properly trained to accurately classify real rare case vs. rare case false alarms.
  • Data-C With respect to unlabelled image data (Data-C), this data is related to a much larger corpus of data compared to original data (Data-A). Preferably, it is required for Data-C to have at least some visually similar samples to those in Data-A. Typically Data-C is 100x or more than Data-A so that the chance of using only portion of unlabelled data that corresponds to specific class is extremely small and therefore can be ignored.
  • Figure 7 represents Stage F of secondary combined feature/decision space in a secondary combined deep neural network (DNN-F).
  • Supplied data are unlabelled images (Data-C) and assigned tags (Label-C), composed of automatically labelled data on the stage E and a penalty matrix.
  • the algorithm trains DNN-F by:
  • Stage F output is secondary combined feature/decision spaces constructed in favour of rare case samples with high penalty, according to the penalty-matrix defined offline.
  • Stage F goal is to create feature/decision space that hosts best both frequent and rare cases with high risk-level of false alarms, according to prior knowledge, i.e. the penalty matrix.
  • the modified loss function step an example could be as follows: since the DNN-F should classify for each input image its category mn out of M * N categories, the cost of misclassification would be 0 if the classification is correct and 1 * (Penalty) if the classification is incorrect. Therefore DNN-F would be penalized for non-recognizing object's class 1 :N and at the same time not estimating location of the most discriminative features for this object.
  • N * N matrix where N are classes from Label-A.
  • N are classes from Label-A.
  • Each value represents normalized application risk associated with misclassification between specific pair of classes defined by row and column.
  • Figure 8 represents Stage G of training "original" feature/decision spaces hosted in a deep convolutional neural network (DNN-G). Supplied data are images (Data-A) and tags (Label-A) and a penalty matrix.
  • Data-A images
  • Label-A tags
  • penalty matrix a penalty matrix
  • the algorithm trains DNN-G by:
  • Stage G output is a primary combined feature/decision space, which is warped from pre-trained secondary space with much larger capacity M * N, constructed in favour of rare case samples with high penalty, according to the penalty-matrix defined offline.
  • Stage G goal is the same as on stage F, with input and output compatible with original data flow of DNN-A to close the reinforcement loop.
  • Stage G since we made sequences of knowledge transfers resulted in our ability to classify rare cases at stage F according to priors defined in penalty matrix, we preferably want to stack one fully connected layer at the end of DNN-F, so that this layer re-routes classification results of DNN-F 1 :M * N into initial class space of 1 :N. More generally, regarding the whole DNN, at each stage, where DNN is train based on previous knowledge, new knowledge is acquired and further transferred along the chain. With such procedure we are improving feature and decision spaces by collecting more samples near rare cases and far from previously learned samples.
  • a vehicle 100 is equipped with a path capturing unit (200; 210) arranged to capture and convert portions of a followed path seen at least from a driver's point of view into a series of digital files, when the vehicle is driven.
  • path capturing unit can be a camera 200 and/or a 360° scanning unit 210, such as a laser light scanning unit (LIDAR), pointing the road ahead to take a video or a continuous series of pictures during a journey.
  • LIDAR laser light scanning unit
  • the vehicle 100 also comprises :
  • a processing unit hosting a deep neural network, arranged to classify both generic and rare cases based on the series of digital files according the image processing method of the invention
  • a display unit arranged to display an information related to the classified generic and rare cases
  • an autonomous driving unit arranged to control the vehicle
  • a decision unit arranged to activate at least one of the display unit and the autonomous driving unit depending on the classified rare cases.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Software Systems (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Medical Informatics (AREA)
  • Aviation & Aerospace Engineering (AREA)
  • Radar, Positioning & Navigation (AREA)
  • Remote Sensing (AREA)
  • Automation & Control Theory (AREA)
  • Business, Economics & Management (AREA)
  • Game Theory and Decision Science (AREA)
  • Electromagnetism (AREA)
  • Multimedia (AREA)
  • Probability & Statistics with Applications (AREA)
  • Image Analysis (AREA)
  • Traffic Control Systems (AREA)

Abstract

L'invention concerne un procédé pour renforcer la capacité d'apprentissage d'un réseau neuronal profond afin de classifier des cas rares, ledit procédé comprend les étapes suivantes : formation d'un premier réseau neuronal profond (DNN-A) utilisé pour classifier des cas génériques de données originales (Données-A) en étiquettes spécifiées (Étiquette-A); localisation de caractéristiques discriminatives spécifiques à la classe à l'intérieur des données originales traitées par le biais du DNN-A et mise en correspondance des caractéristiques discriminative spécifiques à la classe sous la forme d'étiquettes probabilistes spatiales (Étiquette-B); formation d'un deuxième réseau neuronal profond (DNN-C) utilisé pour classifier des cas rares des données originales dans les étiquettes probabilistes spatiales; et formation d'un réseau neuronal profond combiné (DNN-D) utilisé pour classifier à la fois les cas rares et génériques des données originales en étiquettes principales combinées spécifiées et probabilistes spatiales (Étiquette-A+B*).
PCT/EP2017/056172 2016-03-15 2017-03-15 Procédé de classification de cas uniques/rares par renforcement de l'apprentissage dans des réseaux neuronaux WO2017158058A1 (fr)

Priority Applications (3)

Application Number Priority Date Filing Date Title
EP17710934.5A EP3430576A1 (fr) 2016-03-15 2017-03-15 Procédé de classification de cas uniques/rares par renforcement de l'apprentissage dans des réseaux neuronaux
JP2018549257A JP6983800B2 (ja) 2016-03-15 2017-03-15 ニューラルネットワークにおける強化学習によるユニーク／希少なケースの分類のための方法
US16/085,989 US10776664B2 (en) 2016-03-15 2017-03-15 Method for classification of unique/rare cases by reinforcement learning in neural networks

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
DE102016204275.0 2016-03-15
DE102016204275 2016-03-15

Publications (1)

Publication Number Publication Date
WO2017158058A1 true WO2017158058A1 (fr) 2017-09-21

Family

ID=58314237

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/EP2017/056172 WO2017158058A1 (fr) 2016-03-15 2017-03-15 Procédé de classification de cas uniques/rares par renforcement de l'apprentissage dans des réseaux neuronaux

Country Status (4)

Country Link
US (1) US10776664B2 (fr)
EP (1) EP3430576A1 (fr)
JP (1) JP6983800B2 (fr)
WO (1) WO2017158058A1 (fr)

Cited By (33)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108133238A (zh) * 2017-12-29 2018-06-08 国信优易数据有限公司 一种人脸识别模型训练方法和装置及人脸识别方法和装置
CN108197670A (zh) * 2018-01-31 2018-06-22 国信优易数据有限公司 伪标签生成模型训练方法、装置及伪标签生成方法及装置
CN108596078A (zh) * 2018-04-20 2018-09-28 南京信息工程大学 一种基于深度神经网络的海洋噪声信号识别方法
CN108648187A (zh) * 2018-05-14 2018-10-12 南方医科大学 一种基于深度特征袋的分类方法
CN109085751A (zh) * 2018-09-16 2018-12-25 南京大学 一种基于多粒度强化学习的六足机器人导航方法
CN109325547A (zh) * 2018-10-23 2019-02-12 苏州科达科技股份有限公司 非机动车图像多标签分类方法、系统、设备及存储介质
WO2019075771A1 (fr) * 2017-10-20 2019-04-25 Huawei Technologies Co., Ltd. Procédé et système d'auto-apprentissage d'apprentissage semi-supervisé au moyen de réseaux antagonistes génératifs
CN109858557A (zh) * 2019-02-13 2019-06-07 安徽大学 一种新的高光谱图像数据半监督分类算法
CN109885091A (zh) * 2019-03-21 2019-06-14 华北电力大学（保定） 一种无人机自主飞行控制方法及系统
US10346724B2 (en) * 2017-06-22 2019-07-09 Waymo Llc Rare instance classifiers
CN110097033A (zh) * 2019-05-15 2019-08-06 成都电科智达科技有限公司 一种基于特征扩充的单样本人脸识别方法
JP2019153277A (ja) * 2017-12-07 2019-09-12 イムラ ウーロプ ソシエテ・パ・アクシオンス・シンプリフィエ エンドツーエンド深層ニューラルネットワークを使用する危険ランク付け
CN110234085A (zh) * 2019-05-23 2019-09-13 深圳大学 基于对抗迁移网络的室内位置指纹地图生成方法及系统
JP2019164618A (ja) * 2018-03-20 2019-09-26 株式会社東芝 信号処理装置、信号処理方法およびプログラム
CN110287191A (zh) * 2019-06-25 2019-09-27 北京明略软件系统有限公司 数据对齐方法及装置、存储介质、电子装置
CN110659692A (zh) * 2019-09-26 2020-01-07 重庆大学 基于强化学习和深度神经网络的病理图像自动标注方法
JP2020014042A (ja) * 2018-07-13 2020-01-23 日本放送協会 画質評価装置、学習装置及びプログラム
CN110738221A (zh) * 2018-07-18 2020-01-31 华为技术有限公司 一种运算系统及方法
WO2020166303A1 (fr) * 2019-02-15 2020-08-20 オムロン株式会社 Appareil, procédé et programme de génération de modèle, et appareil de prédiction
US10762399B2 (en) 2017-12-18 2020-09-01 Ford Global Technologies, Llc Using deep video frame prediction for training a controller of an autonomous vehicle
WO2020219971A1 (fr) * 2019-04-25 2020-10-29 Google Llc Apprentissage de modèles d'apprentissage machine à l'aide d'une augmentation de données non supervisée
CN111950638A (zh) * 2020-08-14 2020-11-17 厦门美图之家科技有限公司 基于模型蒸馏的图像分类方法、装置和电子设备
US10860878B2 (en) 2019-02-16 2020-12-08 Wipro Limited Method and system for synthesizing three-dimensional data
JP2021501414A (ja) * 2017-11-01 2021-01-14 インターナショナル・ビジネス・マシーンズ・コーポレーションＩｎｔｅｒｎａｔｉｏｎａｌ Ｂｕｓｉｎｅｓｓ Ｍａｃｈｉｎｅｓ Ｃｏｒｐｏｒａｔｉｏｎ 欺瞞勾配の使用を介した、勾配ベースの攻撃からの認知システムの保護
JPWO2019150813A1 (ja) * 2018-01-30 2021-01-28 富士フイルム株式会社 データ処理装置及び方法、認識装置、学習データ保存装置、機械学習装置並びにプログラム
CN112488170A (zh) * 2020-11-24 2021-03-12 杭州电子科技大学 一种基于深度学习的多特征融合图像分类方法
CN112712094A (zh) * 2019-10-24 2021-04-27 北京四维图新科技股份有限公司 模型的训练方法、装置、设备及存储介质
JP2021524103A (ja) * 2018-05-18 2021-09-09 オ−ディーディー コンセプツ インク． 画像内のオブジェクトの代表特性を抽出する方法、装置及びコンピュータプログラム
CN113496481A (zh) * 2021-05-20 2021-10-12 北京交通大学 一种少样本胸部X-Ray图像的辅助检测方法
CN113535953A (zh) * 2021-07-15 2021-10-22 湖南大学 一种基于元学习的少样本分类方法
US11176429B2 (en) 2019-05-13 2021-11-16 International Business Machines Corporation Counter rare training date for artificial intelligence
US20220383111A1 (en) * 2019-09-27 2022-12-01 D5Ai Llc Selective training of deep learning modules
US11853901B2 (en) 2019-07-26 2023-12-26 Samsung Electronics Co., Ltd. Learning method of AI model and electronic apparatus

Families Citing this family (29)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2018060268A (ja) * 2016-10-03 2018-04-12 株式会社日立製作所 認識装置および学習システム
EP3607492A4 (fr) * 2017-04-07 2021-01-06 INTEL Corporation Procédés et systèmes de formation avancée et augmentée de réseaux neuronaux profonds à l'aide de données synthétiques et de réseaux génératifs innovants
JP6729516B2 (ja) * 2017-07-27 2020-07-22 トヨタ自動車株式会社 識別装置
DE102017219274A1 (de) * 2017-10-26 2019-05-02 Robert Bosch Gmbh Verfahren und Vorrichtung zum Verbessern der Robustheit eines maschinellen Lernsystems
US10726313B2 (en) * 2018-04-19 2020-07-28 Adobe Inc. Active learning method for temporal action localization in untrimmed videos
US10990096B2 (en) * 2018-04-27 2021-04-27 Honda Motor Co., Ltd. Reinforcement learning on autonomous vehicles
US11481416B2 (en) 2018-07-12 2022-10-25 International Business Machines Corporation Question Answering using trained generative adversarial network based modeling of text
US11281976B2 (en) * 2018-07-12 2022-03-22 International Business Machines Corporation Generative adversarial network based modeling of text for natural language processing
US10901416B2 (en) * 2018-07-19 2021-01-26 Honda Motor Co., Ltd. Scene creation system for autonomous vehicles and methods thereof
US10832003B2 (en) * 2018-08-26 2020-11-10 CloudMinds Technology, Inc. Method and system for intent classification
CN111178491A (zh) * 2018-11-09 2020-05-19 佳能株式会社 神经网络模型的训练和应用方法、装置、系统及存储介质
US11375176B2 (en) * 2019-02-05 2022-06-28 Nvidia Corporation Few-shot viewpoint estimation
FR3095878B1 (fr) * 2019-05-10 2021-10-08 Univ De Brest Procédé d'analyse automatique d'images pour reconnaître automatiquement au moins une caractéristique rare
CN110180186B (zh) * 2019-05-28 2022-08-19 北京奇思妙想信息技术有限公司 一种地形图转换方法及系统
US11308365B2 (en) 2019-06-13 2022-04-19 Expedia, Inc. Image classification system
US11507090B2 (en) * 2019-06-19 2022-11-22 Uber Technologies, Inc. Systems and methods for vehicle motion control with interactive object annotation
CA3085033C (fr) * 2019-07-30 2023-01-03 Imrsv Data Labs Inc. Methodes et systemes de classification de donnees texte a multiples etiquettes
US20210049473A1 (en) * 2019-08-14 2021-02-18 The Board Of Trustees Of The Leland Stanford Junior University Systems and Methods for Robust Federated Training of Neural Networks
CN112633459A (zh) * 2019-09-24 2021-04-09 华为技术有限公司 训练神经网络的方法、数据处理方法和相关装置
CN110717600B (zh) * 2019-09-30 2021-01-26 京东城市(北京)数字科技有限公司 样本池构建方法和装置、以及算法训练方法和装置
US20210241101A1 (en) * 2020-02-04 2021-08-05 Dsp Group Ltd. Post-training Control of the Bias of Neural Networks
CN111382791B (zh) * 2020-03-07 2023-12-26 北京迈格威科技有限公司 深度学习任务处理方法、图像识别任务处理方法和装置
US10963792B1 (en) * 2020-03-26 2021-03-30 StradVision, Inc. Method for training deep learning network based on artificial intelligence and learning device using the same
US11087883B1 (en) * 2020-04-02 2021-08-10 Blue Eye Soft, Inc. Systems and methods for transfer-to-transfer learning-based training of a machine learning model for detecting medical conditions
CN111461232A (zh) * 2020-04-02 2020-07-28 大连海事大学 一种基于多策略批量式主动学习的核磁共振图像分类方法
CN111723780B (zh) * 2020-07-22 2023-04-18 浙江大学 基于高分辨率遥感图像的跨域数据的定向迁移方法及系统
SG10202007348TA (en) * 2020-08-01 2021-03-30 Sensetime Int Pte Ltd Target object identification method and apparatus
CN112529178B (zh) * 2020-12-09 2024-04-09 中国科学院国家空间科学中心 一种适用于无预选框检测模型的知识蒸馏方法及系统
CN113327666B (zh) * 2021-06-21 2022-08-12 青岛科技大学 一种胸片疾病多分类网络的多标签局部至全局学习方法

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7610130B1 (en) * 2005-07-20 2009-10-27 Sandia Corporation Physical context management for a motor vehicle
US7860813B2 (en) * 2008-01-24 2010-12-28 Nec Laboratories America, Inc. Monitoring driving safety using semi-supervised sequential learning
JP2012026982A (ja) * 2010-07-27 2012-02-09 Panasonic Electric Works Sunx Co Ltd 検査装置
CA2944829C (fr) * 2014-05-23 2022-10-25 Ting Chen Systemes et procedes de detection de structures et/ou de motifs biologiques dans des images

Non-Patent Citations (10)

* Cited by examiner, † Cited by third party
Title
ADRIANA ROMERO ET AL: "FitNets: hints for thin deep nets", ARXIV:1412.6550V4, 27 March 2015 (2015-03-27), XP055349753, Retrieved from the Internet <URL:https://arxiv.org/abs/1412.6550v4> [retrieved on 20170227] *
CHEN XINLEI ET AL: "Webly Supervised Learning of Convolutional Networks", 2015 IEEE INTERNATIONAL CONFERENCE ON COMPUTER VISION (ICCV), IEEE, 7 December 2015 (2015-12-07), pages 1431 - 1439, XP032866473, DOI: 10.1109/ICCV.2015.168 *
DAVID LOPEZ-PAZ ET AL: "Unifying distillation and privileged information", 26 February 2016 (2016-02-26), XP055377961, Retrieved from the Internet <URL:https://arxiv.org/pdf/1511.03643.pdf> [retrieved on 20170601] *
GEOFFREY HINTON; ORIOL VINYALS; JEFF DEAN: "Distilling the Knowledge in a Neural Network", ARXIV:1503.02531
IAN J. GOODFELLOW; JEAN POUGET-ABADIE; MEHDI MIRZA; BING XU; DAVID WARDE-FARLEY; SHERJIL OZAIR; AARON COURVILLE; YOSHUA BENGIO: "Generative Adversarial Networks", ARXIV:1406.2661
IAN J. GOODFELLOW; JONATHON SHLENS; CHRISTIAN SZEGEDY: "Explaining and Harnessing Adversarial Examples", ARXIV:1412.6572
MINGSHENG LONG ET AL: "Learning Multiple Tasks with Deep Relationship Networks", 6 June 2015 (2015-06-06), XP055378244, Retrieved from the Internet <URL:https://arxiv.org/pdf/1506.02117v1.pdf> [retrieved on 20170601] *
OQUAB MAXIME ET AL: "Learning and Transferring Mid-level Image Representations Using Convolutional Neural Networks", 2014 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION, IEEE, 23 June 2014 (2014-06-23), pages 1717 - 1724, XP032649359, DOI: 10.1109/CVPR.2014.222 *
SANGHEUM HWANG ET AL: "Self-Transfer Learning for Fully Weakly Supervised Object Localization", ARXIV.ORG, CORNELL UNIVERSITY LIBRARY, 201 OLIN LIBRARY CORNELL UNIVERSITY ITHACA, NY 14853, 4 February 2016 (2016-02-04), XP080681467 *
SHUMEET BALUJA MICHELE COVELL; RAHUL SUKTHANKAR: "The Virtues of Peer Pressure: A Simple Method for Discovering High-Value Mistakes", CAIP, 2015

Cited By (53)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10346724B2 (en) * 2017-06-22 2019-07-09 Waymo Llc Rare instance classifiers
US11120337B2 (en) 2017-10-20 2021-09-14 Huawei Technologies Co., Ltd. Self-training method and system for semi-supervised learning with generative adversarial networks
WO2019075771A1 (fr) * 2017-10-20 2019-04-25 Huawei Technologies Co., Ltd. Procédé et système d'auto-apprentissage d'apprentissage semi-supervisé au moyen de réseaux antagonistes génératifs
JP7059368B2 (ja) 2017-11-01 2022-04-25 インターナショナル・ビジネス・マシーンズ・コーポレーション 欺瞞勾配の使用を介した、勾配ベースの攻撃からの認知システムの保護
JP2021501414A (ja) * 2017-11-01 2021-01-14 インターナショナル・ビジネス・マシーンズ・コーポレーションＩｎｔｅｒｎａｔｉｏｎａｌ Ｂｕｓｉｎｅｓｓ Ｍａｃｈｉｎｅｓ Ｃｏｒｐｏｒａｔｉｏｎ 欺瞞勾配の使用を介した、勾配ベースの攻撃からの認知システムの保護
JP2019153277A (ja) * 2017-12-07 2019-09-12 イムラ ウーロプ ソシエテ・パ・アクシオンス・シンプリフィエ エンドツーエンド深層ニューラルネットワークを使用する危険ランク付け
JP7217138B2 (ja) 2017-12-07 2023-02-02 イムラ ウーロプ ソシエテ・パ・アクシオンス・シンプリフィエ エンドツーエンド深層ニューラルネットワークを使用する危険ランク付け
US10762399B2 (en) 2017-12-18 2020-09-01 Ford Global Technologies, Llc Using deep video frame prediction for training a controller of an autonomous vehicle
CN108133238A (zh) * 2017-12-29 2018-06-08 国信优易数据有限公司 一种人脸识别模型训练方法和装置及人脸识别方法和装置
CN108133238B (zh) * 2017-12-29 2020-05-19 国信优易数据有限公司 一种人脸识别模型训练方法和装置及人脸识别方法和装置
US11288550B2 (en) 2018-01-30 2022-03-29 Fujifilm Corporation Data processing apparatus and method, recognition apparatus, learning data storage apparatus, machine learning apparatus, and program
JPWO2019150813A1 (ja) * 2018-01-30 2021-01-28 富士フイルム株式会社 データ処理装置及び方法、認識装置、学習データ保存装置、機械学習装置並びにプログラム
CN108197670A (zh) * 2018-01-31 2018-06-22 国信优易数据有限公司 伪标签生成模型训练方法、装置及伪标签生成方法及装置
CN108197670B (zh) * 2018-01-31 2021-06-15 国信优易数据股份有限公司 伪标签生成模型训练方法、装置及伪标签生成方法及装置
JP2019164618A (ja) * 2018-03-20 2019-09-26 株式会社東芝 信号処理装置、信号処理方法およびプログラム
CN108596078A (zh) * 2018-04-20 2018-09-28 南京信息工程大学 一种基于深度神经网络的海洋噪声信号识别方法
CN108648187B (zh) * 2018-05-14 2021-06-15 南方医科大学 一种基于深度特征袋的分类方法
CN108648187A (zh) * 2018-05-14 2018-10-12 南方医科大学 一种基于深度特征袋的分类方法
JP2021524103A (ja) * 2018-05-18 2021-09-09 オ−ディーディー コンセプツ インク． 画像内のオブジェクトの代表特性を抽出する方法、装置及びコンピュータプログラム
JP7202091B2 (ja) 2018-07-13 2023-01-11 日本放送協会 画質評価装置、学習装置及びプログラム
JP2020014042A (ja) * 2018-07-13 2020-01-23 日本放送協会 画質評価装置、学習装置及びプログラム
CN110738221A (zh) * 2018-07-18 2020-01-31 华为技术有限公司 一种运算系统及方法
CN110738221B (zh) * 2018-07-18 2024-04-26 华为技术有限公司 一种运算系统及方法
CN109085751A (zh) * 2018-09-16 2018-12-25 南京大学 一种基于多粒度强化学习的六足机器人导航方法
CN109325547A (zh) * 2018-10-23 2019-02-12 苏州科达科技股份有限公司 非机动车图像多标签分类方法、系统、设备及存储介质
CN109858557A (zh) * 2019-02-13 2019-06-07 安徽大学 一种新的高光谱图像数据半监督分类算法
CN109858557B (zh) * 2019-02-13 2022-12-27 安徽大学 一种新的高光谱图像数据半监督分类方法
WO2020166303A1 (fr) * 2019-02-15 2020-08-20 オムロン株式会社 Appareil, procédé et programme de génération de modèle, et appareil de prédiction
JP2020135243A (ja) * 2019-02-15 2020-08-31 オムロン株式会社 モデル生成装置、予測装置、モデル生成方法、及びモデル生成プログラム
JP7162550B2 (ja) 2019-02-15 2022-10-28 オムロン株式会社 モデル生成装置、予測装置、モデル生成方法、及びモデル生成プログラム
US10860878B2 (en) 2019-02-16 2020-12-08 Wipro Limited Method and system for synthesizing three-dimensional data
CN109885091A (zh) * 2019-03-21 2019-06-14 华北电力大学（保定） 一种无人机自主飞行控制方法及系统
WO2020219971A1 (fr) * 2019-04-25 2020-10-29 Google Llc Apprentissage de modèles d'apprentissage machine à l'aide d'une augmentation de données non supervisée
US11176429B2 (en) 2019-05-13 2021-11-16 International Business Machines Corporation Counter rare training date for artificial intelligence
CN110097033B (zh) * 2019-05-15 2023-04-07 成都电科智达科技有限公司 一种基于特征扩充的单样本人脸识别方法
CN110097033A (zh) * 2019-05-15 2019-08-06 成都电科智达科技有限公司 一种基于特征扩充的单样本人脸识别方法
CN110234085A (zh) * 2019-05-23 2019-09-13 深圳大学 基于对抗迁移网络的室内位置指纹地图生成方法及系统
CN110234085B (zh) * 2019-05-23 2020-09-15 深圳大学 基于对抗迁移网络的室内位置指纹地图生成方法及系统
CN110287191B (zh) * 2019-06-25 2021-07-27 北京明略软件系统有限公司 数据对齐方法及装置、存储介质、电子装置
CN110287191A (zh) * 2019-06-25 2019-09-27 北京明略软件系统有限公司 数据对齐方法及装置、存储介质、电子装置
US11853901B2 (en) 2019-07-26 2023-12-26 Samsung Electronics Co., Ltd. Learning method of AI model and electronic apparatus
CN110659692A (zh) * 2019-09-26 2020-01-07 重庆大学 基于强化学习和深度神经网络的病理图像自动标注方法
CN110659692B (zh) * 2019-09-26 2023-04-18 重庆大学 基于强化学习和深度神经网络的病理图像自动标注方法
US20220383111A1 (en) * 2019-09-27 2022-12-01 D5Ai Llc Selective training of deep learning modules
CN112712094A (zh) * 2019-10-24 2021-04-27 北京四维图新科技股份有限公司 模型的训练方法、装置、设备及存储介质
CN111950638B (zh) * 2020-08-14 2024-02-06 厦门美图之家科技有限公司 基于模型蒸馏的图像分类方法、装置和电子设备
CN111950638A (zh) * 2020-08-14 2020-11-17 厦门美图之家科技有限公司 基于模型蒸馏的图像分类方法、装置和电子设备
CN112488170A (zh) * 2020-11-24 2021-03-12 杭州电子科技大学 一种基于深度学习的多特征融合图像分类方法
CN112488170B (zh) * 2020-11-24 2024-04-05 杭州电子科技大学 一种基于深度学习的多特征融合图像分类方法
CN113496481A (zh) * 2021-05-20 2021-10-12 北京交通大学 一种少样本胸部X-Ray图像的辅助检测方法
CN113496481B (zh) * 2021-05-20 2023-11-07 北京交通大学 一种少样本胸部X-Ray图像的辅助检测方法
CN113535953A (zh) * 2021-07-15 2021-10-22 湖南大学 一种基于元学习的少样本分类方法
CN113535953B (zh) * 2021-07-15 2022-05-27 湖南大学 一种基于元学习的少样本分类方法

Also Published As

Publication number Publication date
US10776664B2 (en) 2020-09-15
JP6983800B2 (ja) 2021-12-17
JP2019514107A (ja) 2019-05-30
US20190122077A1 (en) 2019-04-25
EP3430576A1 (fr) 2019-01-23

Similar Documents

Publication Publication Date Title
US10776664B2 (en) Method for classification of unique/rare cases by reinforcement learning in neural networks
US10663594B2 (en) Processing method of a 3D point cloud
CN107851191B (zh) 用于图像中的对象检测的基于上下文的先验
JP2020520007A (ja) ネットワーク計算エッジにわたってアプリケーションを連続して操作する、人工知能および深層学習においてメモリで境界された継続学習を可能にするシステムおよび方法
US20240095972A1 (en) Image processing apparatus and method for style transformation
Boualia et al. Pose-based human activity recognition: a review
Wong et al. Track everything: Limiting prior knowledge in online multi-object recognition
US20230099521A1 (en) 3d map and method for generating a 3d map via temporal and unified panoptic segmentation
Liu et al. Deep transfer learning for intelligent vehicle perception: A survey
Saleh et al. Cyclist intent prediction using 3d lidar sensors for fully automated vehicles
Silva et al. Multi-view fine-grained vehicle classification with multi-loss learning
EP3767534A1 (fr) Dispositif et procédé d&#39;évaluation d&#39;un déterminateur de carte de relief
US20230281961A1 (en) System and method for 3d object detection using multi-resolution features recovery using panoptic segmentation information
US20210390419A1 (en) Device and Method for Training and Testing a Classifier
Sjöberg et al. Indoor location recognition using fusion of SVM-based visual classifiers
EP3629237B1 (fr) Dispositif et procédé pour améliorer la robustesse face à des exemples contradictoires
Unar et al. Towards applying image retrieval approach for finding semantic locations in autonomous vehicles
Miller Epistemic uncertainty estimation for object detection in open-set conditions
Zhang et al. A Viewpoint Adaptation Ensemble Contrastive Learning framework for vessel type recognition with limited data
Nobis Autonomous driving: Radar sensor noise filtering and multimodal sensor fusion for object detection with artificial neural net-works
Tasnim et al. Progress in Object Detection: An In-Depth Analysis of Methods and Use Cases
Kozel et al. Real-Time Traffic Lights Identification using YOLOv3 Algorithm For Autonomous Vehicles
HAMADI Real Time Multi-Object Tracking Using Deep Learning
Weill Edge-Computing Deep Learning-Based Computer Vision Systems
US20230326005A1 (en) Data augmentation for domain generalization

Legal Events

Date Code Title Description
ENP Entry into the national phase

Ref document number: 2018549257

Country of ref document: JP

Kind code of ref document: A

NENP Non-entry into the national phase

Ref country code: DE

WWE Wipo information: entry into national phase

Ref document number: 2017710934

Country of ref document: EP

ENP Entry into the national phase

Ref document number: 2017710934

Country of ref document: EP

Effective date: 20181015

121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 17710934

Country of ref document: EP

Kind code of ref document: A1